ZFS performance tuning tips

Size must be power of 2.

• ZFS file system: recordsize (default 128kB)
• ZVOL block device: volblocksize (default - Solaris based - was 8kB). With OpenZFS 2.2 default was changed to 16k to “reduce wasted space”.

This is basic operation unit for ZFS. ZFS is COW filesystem. So to modify even one byte of data stored inside 128kB record it must read 128kB record, modify it and store 128kB in new place. It creates huge read and write amplification.

Small sizes:

• are good for dedicated workloads, like databases, etc.
• 4kB has no sense with compression. Even if data is compressed it still takes smaller unit - 4kB.
• slower sequential read - lots of IOPS and checksum checks
• fragmentation over time
• metadata overhead (4kB data size needs also metada, so another 4kB unit will be used).

Big size:

• benefit from compression. I.e: 128kB data compressed to 16kB will create 16kB record size.
• very good for storage (write once data)
• read/write amplification for small read/writes
• good for sequential access
• good for HDDs (less fragmentation = less seeks)
• less metadata
• less fragmentation

Note: recordsize / volblocksize only defines upper limit. Smaller data still can create smaller recordsize (is it true for block?).

Examples:

• 16kB for MySQL/InnoDB
• 128kB for rotational HDDs

When huge portion of data are written (new backups) or read (backup verify) L2ARC is constantly written with current data. To change this behaviour to cache only Most Frequent Use:

/etc/modprobe.d/zfs.conf

options zfs l2arc_mfuonly=1 l2arc_noprefetch=0

Explanation:

• l2arc_mfuonly Controls whether only MFU metadata and data are cached from ARC into L2ARC. This may be desirable to avoid wasting space on L2ARC when reading/writing large amounts of data that are not expected to be accessed more than once. By default both MRU and MFU data and metadata are cached in the L2ARC.
• l2arc_noprefetch Disables writing prefetched, but unused, buffers to cache devices. Setting to 0 can increase L2ARC hit rates for workloads where the ARC is too small for a read workload that benefits from prefetching. Also, if the main pool devices are very slow, setting to 0 can improve some workloads such as backups.

If whole device is managed by ZFS (not partition), ZFS sets scheduler to none.

For rotational devices, there is no sense to use advanced schedulers cfq or bfq directly on hard disc. Both depends on processes, processes groups and application. In this case there is group of kernel processess for ZFS.

Only possible scheduler to consider is deadline / mq-deadline. Deadline scheduler group reads into batches and writed into separate batches ordering by increasing LBA address (so it should be good for HDDs).

There is a discussion on OpenZFS project to do not touch schedulers anymore and let it to be configured by admin:

• Set "none" scheduler if available (initramfs) #9042
• https://github.com/openzfs/zfs/commit/42c24d90d112b6e9e1a304346a1335e058f1678b

There is huge benefit to use bfq on rotational HDD. No more huge lags during KVM backups.

bfq honor ionice and:

• kernel zvol processes have prio be/0
• kvm processes have prio be/4
• kvm process during vzdump have be/7 - NOTE: only with patched version of kvm: pve-qemu-kvm >= 8.1.5-6.

ZFS Send & RaidZ - Poor performance on HDD #14916

cat /sys/module/zfs/parameters/zfs_vdev_async_write_max_active 
echo 2 > /sys/module/zfs/parameters/zfs_vdev_async_write_max_active 
cat /sys/module/zfs/parameters/zfs_vdev_async_write_max_active

Use huge record size - it can help on SMR drives. Note: it only make sense for ZFS file system. Cannot be applied on ZVOL.

zfs set recordsize=1M hddpool/data
zfs set recordsize=1M hddpool/vz

NOTE: SMR drives behaves correctly for sequential writes, but long working ZFS or LVM thin spread writes into lots of random location causing unusable IOPS. So never use SMR.

For ZVOLs: zvol volblocksize

Change ZFS volblock on a running Proxmox VM

Note: When no stripping is used (simple mirror) volblocksize should be 4kB (or at least the same as ashift).

Note: Latest Proxmox default volblock size was increased form 8k to 16k. When 8k is used warning is shown:

Warning: volblocksize (8192) is less than the default minimum block size (16384).
To reduce wasted space a volblocksize of 16384 is recommended.

zfs create -s -V 40G hddpool/data/vm-156-disk-0-16k -o volblock=16k
dd if=/dev/zvol/hddpool/data/vm-156-disk-0 of=/dev/zvol/hddpool/data/vm-156-disk-0-16k bs=1M status=progress conv=sparse
zfs rename hddpool/data/vm-156-disk-0 hddpool/data/vm-156-disk-0-backup
zfs rename hddpool/data/vm-156-disk-0-16k hddpool/data/vm-156-disk-0

Use bfq is mandator. See my findings.

See Archlinux wiki: Databases

zfs set recordsize=8K <pool>/postgres
 
# ONLY for SSD/NVM devices:
zfs set logbias=throughput <pool>/postgres

By default ZFS can sue 50% of RAM for ARC cache:

# apt install zfsutils-linux
 
# arcstat 
    time  read  miss  miss%  dmis  dm%  pmis  pm%  mmis  mm%  size     c  avail
16:47:26     3     0      0     0    0     0    0     0    0   15G   15G   1.8G

# arc_summary 
 
ARC size (current):                                    98.9 %   15.5 GiB
        Target size (adaptive):                       100.0 %   15.6 GiB
        Min size (hard limit):                          6.2 %  999.6 MiB
        Max size (high water):                           16:1   15.6 GiB
        Most Frequently Used (MFU) cache size:         75.5 %   11.2 GiB
        Most Recently Used (MRU) cache size:           24.5 %    3.6 GiB
        Metadata cache size (hard limit):              75.0 %   11.7 GiB
        Metadata cache size (current):                  8.9 %    1.0 GiB
        Dnode cache size (hard limit):                 10.0 %    1.2 GiB
        Dnode cache size (current):                     5.3 %   63.7 MiB

ARC size can be tuned by settings zfs kernel module parameters (Module Parameters):

• zfs_arc_max: Maximum size of ARC in bytes. If set to 0 then the maximum size of ARC is determined by the amount of system memory installed (50% on Linux)
• zfs_arc_min: Minimum ARC size limit. When the ARC is asked to shrink, it will stop shrinking at c_min as tuned by zfs_arc_min.
• zfs_arc_meta_limit_percent: Sets the limit to ARC metadata, arc_meta_limit, as a percentage of the maximum size target of the ARC, c_max. Default is 75.

Proxmox recommends following rule:

As a general rule of thumb, allocate at least 2 GiB Base + 1 GiB/TiB-Storage

Set zfs_arc_max to 4GB and zfs_arc_min to 128MB:

echo "$[4 * 1024*1024*1024]" >/sys/module/zfs/parameters/zfs_arc_max
echo "$[128     *1024*1024]" >/sys/module/zfs/parameters/zfs_arc_min

Make options persistent:

options zfs zfs_prefetch_disable=1
options zfs zfs_arc_max=4294967296
options zfs zfs_arc_min=134217728
options zfs zfs_arc_meta_limit_percent=75

and update-initramfs -u